Physics-prior and deep learning fusion for single-plane diffractive imaging in frequency-spatial domains.

Area of Science:

• Optics and Photonics
• Computational Imaging
• Optical Engineering

Background:

• Traditional optical systems are bulky and complex, requiring multiple lenses for high-quality imaging.
• Single-plane diffractive optical elements (SPDOEs) offer a path to compact, single-element imaging systems.
• Existing methods struggle with image reconstruction quality from SPDOEs.

Purpose of the Study:

• To develop a novel prior-guided computational imaging method for high-quality image reconstruction using SPDOEs.
• To design and fabricate a practical SPDOE and analyze its optical characteristics.
• To introduce an attention-enhanced deblurring network tailored for SPDOE imaging challenges.

Main Methods:

• Designed and fabricated an SPDOE (f/5, 50mm focal length).
• Developed a prior-guided attention-enhanced multiscale deblurring network (PAMDN).
• Integrated spatial- and frequency-domain feature extraction and fusion for image reconstruction.

Main Results:

• PAMDN achieved a 30% average improvement in peak signal-to-noise ratio (PSNR) over conventional deconvolution.
• Image quality gains exceeded 13 dB across various field-of-view regions.
• Demonstrated high-quality image reconstruction from a fabricated SPDOE.

Conclusions:

• The proposed PAMDN method significantly enhances image reconstruction quality for SPDOEs.
• This work provides a theoretical basis for high-quality SPDOE imaging.
• Presents a new approach for the miniaturization of optical systems.